TL;DR
This paper presents a computational linguistics approach to deciphering the ancient Proto-Elamite script, a system that has resisted full understanding for over a century. By applying hierarchical clustering and topic modeling to a digital corpus of Proto-Elamite texts, the researchers not only replicate previously known insights but also uncover novel relationships between signs, offering a powerful new toolkit for future decipherment efforts.
What Problem Does This Paper Address?
The Proto-Elamite (PE) script, used in ancient Iran around 3100-2900 BC, is the most widespread writing system from its era and the largest corpus of material in an undeciphered script. While its administrative nature and some basic features like numeric systems are understood, the meaning of most signs and the underlying language remain unknown. Traditional decipherment efforts have been hampered by issues like inaccurate transcriptions and limited access to high-quality images. This paper aims to demonstrate how computational linguistics techniques, applied to a newly available digital corpus, can aid in decipherment by revealing hidden patterns and relationships within the script.
Methods
The researchers employ three primary computational methods:
- Hierarchical Sign Clustering: This technique groups signs based on how frequently they appear together in sequences. They used three different clustering algorithms (neighbor-based, HMM, and generalized Brown clustering) and looked for "stable clusters" that recurred across all methods to identify meaningful groupings. This approach aims to uncover signs that likely function similarly, mirroring how human experts group signs based on context.
- n-Gram Frequency Analysis: This involves counting the occurrences of sequences of signs (unigrams, bigrams, trigrams, etc.). This helps understand the repetitiveness of certain sign combinations, which can indicate common phrases or administrative formulas.
- Latent Dirichlet Allocation (LDA) Topic Modeling: LDA is used to discover abstract "topics" within the corpus, grouping texts that discuss similar subjects. Crucially, LDA can operate on undeciphered texts by analyzing co-occurrence patterns of signs, thus inferring topical relationships without needing to know the meaning of individual signs.
Datasets
The study utilizes the Proto-Elamite (PE) corpus provided by the Cuneiform Digital Library Initiative (CDLI). After removing tablets with only unreadable or numeric signs, the dataset comprises 1399 distinct texts. The working signlist derived from this corpus contains 49 numeric signs and 1623 non-numeric signs, including basic signs and their variants. No new datasets or benchmarks were introduced; the focus was on applying existing computational methods to the established PE corpus.
Key Insights and Ideas
- Computational Linguistics as a Decipherment Aid: The paper strongly advocates for the utility of computational methods in deciphering ancient scripts, demonstrating that they can not only confirm existing hypotheses but also reveal previously unseen connections.
- Stable Clusters Reveal Functional Relationships: The identification of "stable clusters" of signs across multiple clustering algorithms suggests that these signs likely share a functional relationship, even if their precise meaning is unknown. For example, a stable cluster of five signs is hypothesized to represent names syllabically.
- Repetitiveness Varies by Sequence Length: While some sign sequences (bigrams and trigrams) show a degree of repetition, the low frequency of longer n-grams (4-grams and above) suggests that complex linguistic structures might not be heavily encoded in PE, or that decipherment methods relying on extensive repetition might be less effective.
- LDA Uncovers Text Genres and Sign Associations: LDA successfully identified known administrative genres within the PE corpus and, more importantly, highlighted unexpected associations between signs that do not co-occur directly but appear in similar topical contexts. This suggests subtle semantic links that might be missed by manual analysis.
- Sign Variants and Their Significance: The study acknowledges the challenge of determining whether sign variants are semantically distinct or purely graphical, and their approach of treating them as distinct until proven otherwise yields interpretable results.
Key Results
- Replication of Manual Findings: Hierarchical clustering successfully identified known groupings of signs, such as those hypothesized to be used in personal names (e.g., M057, M066, M096, M218, M371).
- Novel Sign Groupings: The analysis revealed new, stable clusters of signs with no previously recognized relationship, such as M003~b and M263~a, which are graphically similar to proto-cuneiform signs related to worker categories.
- Frequency Analysis: The most common unigrams include "object" signs and signs from the hypothesized syllabary. The n-gram analysis shows that even the most frequent trigrams appear in a small percentage of texts, indicating limited repetition for longer sequences.
- LDA Topic Identification: LDA successfully grouped texts into known administrative genres (e.g., labor administration, livestock husbandry). For instance, Topic 1 highlighted the association of M376 (worker/cattle) and M056~f (plow), noting their absence in the same texts but shared topical context. Topic 3 identified a subset of accounts related to M297 ("bread") and M297~b ("keg").
Limitations and Future Work
- Sign Variant Interpretation: The paper notes that a full study of sign variants collapsed together is pending, which might reveal new topics or alter existing ones.
- Model Perplexity and Coherence: While a 10-topic LDA model was used for interpretability, the authors acknowledge that model perplexity drops and coherence peaks at higher topic numbers (around 80 and 110, respectively), suggesting future work could explore larger models.
- Interpreting LDA Topics: The authors provide brief interpretations of some LDA topics, acknowledging that a full analysis of each topic is beyond the scope of the paper.
- Fuzzy String Matching: The paper mentions the development of tools for fuzzy string matching to facilitate further investigation of similar sign strings, which could aid in identifying new sign ordering principles and potential anthroponyms.
- Broader Computational Toolkit: The authors emphasize that the methods used are not exhaustive and that computational approaches can significantly aid progress in a field with limited researchers. They are releasing their data and code to facilitate future discoveries.
Summary updated